diff --git a/design/author.md b/design/author.md index 0591a60..4832f1b 100644 --- a/design/author.md +++ b/design/author.md @@ -15,7 +15,7 @@ already holds text to the content check's rules, from the pack an untrusted chec `apps/skills` plays every example untrusted too. Measured on a five-question quiz, at four characters a token: `content.json` was about 3,470 tokens to write by hand, after reading 3,900 to 9,500 of a type's semantics; the source is 1,000 to 1,800 tokens, written from a type file of 200 -to 3,000. Twenty-eight things the code does not say by itself: +to 3,000. Twenty-nine things the code does not say by itself: - **One generator, three ways to run it.** The third, `saved`, checks content as H5P stores it (*Saved content is checked as the build checks what it writes*, below). `ContentGenerator` was the sweep's @@ -290,10 +290,11 @@ to 3,000. Twenty-eight things the code does not say by itself: the author package pins, so the page plays with the pack it built from. A chat with no tool to run at all, as ChatGPT on a free account once its data-analysis quota is spent (for hours, the owner's test of 2026-11-09), cannot write the link: the skill has it put the whole source in its - reply, since that test's chat gave the page's bare address, which builds nothing. Since - 2026-11-14 a chat writes a page that makes the link, with or without a tool (*In a chat that - built nothing, a page the chat writes makes the link*, below), and a source goes over in a - `json` block only where it names files. + reply, since that test's chat gave the page's bare address, which builds nothing. From + 2026-11-14 a chat wrote a page that made the link (*In a chat that built nothing, a page the chat + writes makes the link*, below), and a source went over in a `json` block only where it named + files. Since 2026-11-15 it writes the link itself, the source in it as JSON, for a source that + names files too (*In route 4 the model writes the link, the source in it as JSON*, below). A link is only what a command printed (2026-11-11). A link ChatGPT gave the owner carried 3,167 characters that were no deflate stream, its table broken at byte 82 of 2,375: the chat had typed them itself, made up or copied wrong, and the page built nothing. So the skill says it @@ -302,6 +303,9 @@ to 3,000. Twenty-eight things the code does not say by itself: block and no link under the old text and the new (12 replies, half of them asked for a link to click), so the rule costs that position nothing, and what it changes in ChatGPT is not measured. No rule reaches a link copied wrong between the command's output and the reply. + This holds for the compressed link of routes 1 to 3. Since 2026-11-15 route 4's link carries + the source as readable JSON, which the model writes as its own text and the build page reads + leniently (*In route 4 the model writes the link, the source in it as JSON*, below). An install that hangs is one that fails (2026-11-13): ChatGPT, asked why it had built nothing, said its `npx` had timed out, and step 3's list named only errors. Haiku in a terminal whose registry took each connection and never answered went on to the two files under the old text @@ -783,7 +787,9 @@ to 3,000. Twenty-eight things the code does not say by itself: 2: built from the archive, made one connection to the registry, and handed the package over unverified, with its files and the link its command printed. With npm working, 1 of 1 ran the verifier as before and handed over its pass. -- **In a chat that built nothing, a page the chat writes makes the link** (2026-11-14, #202). The +- **In a chat that built nothing, a page the chat writes makes the link** (2026-11-14, #202; + replaced on 2026-11-15 by *In route 4 the model writes the link, the source in it as JSON*, + below). The link to the build page was the weak point of a chat that cannot build: compressed text that a command prints and the model copies into its reply. ChatGPT broke 2 of 3 on 2026-11-13 (*The hand-over carries the source and one web address*, above), and with its code tool unavailable, as @@ -890,9 +896,11 @@ to 3,000. Twenty-eight things the code does not say by itself: `two-files.md`), anything else (route 4, `browser.md`). Routes 1 to 3 hand over by `built-here.md`: the build's problems and the file's name, the verifier, run only where npm answered, the link, written in the verifier's command line, and the hand-over. - - Route 4 writes the page; outside a chat, it writes the page to an `.html` file, which makes the - link from a `file:` address as well, where it used to give a link a command printed, so that the - link's commands are written once, in `built-here.md`. + - Route 4 runs no link command: the link's commands are written once, in `built-here.md`. It + wrote the page, and outside a chat the page to an `.html` file, until 2026-11-15. It now hands + over a link the model writes, and outside a chat also saves the source as `.source.json`, which + the build page opens under *Or open a file* (*In route 4 the model writes the link, the source + in it as JSON*, below). What a chat reads of the skill's own text, `SKILL.md` and its route's files: 9.8 KB in the user's browser, 11.3 with npm, 11.5 from the archive and 12.0 from two files, where every route @@ -925,6 +933,49 @@ to 3,000. Twenty-eight things the code does not say by itself: - A run called the Skill tool and got the skill installed on this computer, an older text: the lab's prompts now say the reader is the skill. Haiku also read a route's file before its check now and then, at most 1.5 KB more. +- **In route 4 the model writes the link, the source in it as JSON** (2026-11-15, #206, the + owner's call: "i dont want a code snippet at all here"). The page of *In a chat that built + nothing, a page the chat writes makes the link*, above, worked in ChatGPT on a free account (the + owner, 2026-11-15), but the person saw a block of HTML, opened it with Preview, waited, and clicked + a small link. A model cannot compress a source, but it writes its own JSON, so the build page + reads `#source={…}` too (missing-elements.github.io#54), and route 4 is one Markdown link the model + writes: the source on one line, eight characters escaped (space, `%`, `&`, `\`, `(`, `)`, `<`, + `>`), the rest as it is, letters of any language too. No code block, no command, no Preview. A + source that names files gets the link too, since the page asks for the files before it builds. + Outside a chat the source also goes to a `.source.json` file, since a terminal may not open such + a link. Routes 1 to 3 keep the compressed link their command prints. + + The page reads the readable form leniently, because a weak model's JSON slips (the owner found a + free account's Gemini 3.5 Flash-Lite poor at plain JSON). Each run of escapes is decoded where it + is UTF-8, a bare `%` is kept, a link run on into its sentence is read up to its last `}`, up to + three braces or brackets left open at the end are closed, and text that is still no JSON goes in + the box with where it breaks, for the chat to fix. Only a compressed link that cannot be read is + called damaged. Eleven kinds of link, rendered by markdown-it as a chat renders one and opened in + Chromium, Firefox and WebKit: 33 of 33 as expected. + + Measured on Haiku with a logging skill reader: each reply rendered by markdown-it, its link read + by Node's URL as a browser reads it, decoded as the page decodes it, built, and verified. + + - **Seven escapes, the reply's parts in prose:** 5 of 7 built and passed: Question Sets in + English and Ukrainian, a Drag the Words with `%` and parentheses, a three-chapter Interactive + Book, and a Question Set from a code tool with no network. A True/False left one closing `}` + out. An Accordion on HTML tags wrote `<`, which Markdown turns into `<`, a raw `\`, and no + `)` to close the link, so Markdown made no link of it. + - **So `&` is escaped too, the text says the link ends with a `)` after the last `}`, and the + page closes what is left open:** the True/False then built. With that text, 12 of 12 built + and passed, none needing a closer: three Accordions on HTML tags, three True/False questions + naming `cat.jpg`, five Question Sets (one in Ukrainian, three from a code tool with no + network), and a Drag the Words. No reply held a code block or the source a second time. + - **The archive line,** offered where a code tool runs: "where Node runs here and nothing could be + installed or downloaded" went to 1 of 6 runs with no code tool, wrongly; "where a command you + ran here found Node", to 0 of 1 run with one; "where you have a code tool", as the last + paragraph, to 1 of 2. As an item in a list of what the reply holds, 2 of 2 runs with a code + tool gave it and the run without one did not. + + Links ran from 0.5 to 6.9 KB as written, and a Ukrainian one to 16.6 KB once a renderer escapes + its letters. A route-4 chat read 13 to 35 KB of the skill, most of it type files: the route's file is + 2.1 KB, where the page's was 3.1. Haiku is a floor here: it shows the words are followed, not what a + chat app's renderer does with a link of several KB. That is the owner's run in ChatGPT. - **The skill names no product** (2026-11-12, the owner's call). Its routes were told apart by the apps they were measured in; they now say what each needs: a shell that reaches the npm registry, Node and a download where npm cannot install, or nothing that runs or fetches. What a diff --git a/skills/h5p-author/references/build/browser.md b/skills/h5p-author/references/build/browser.md index b2eadf6..e3bbe28 100644 --- a/skills/h5p-author/references/build/browser.md +++ b/skills/h5p-author/references/build/browser.md @@ -4,52 +4,46 @@ No package is built here. [The build page](https://missing-elements.github.io/h5 it in the user's browser from the source alone, with no install. The user plays it there, which is its only check, and downloads it from there: it is **unverified** until they have. -**Hand the source over in a page that makes the link**, whether or not a command runs here (a -chat's code tool can be unavailable for hours), and whether or not you can tell how the chat -shows it. Copy the page below with the whole source in place of the line `SOURCE`, and every -other line as it is: change nothing in its script. The user's browser makes the link from the -source, so nothing compressed is typed or copied; a link copied into a reply from a command's -output can arrive damaged, so run no command for one. - -- **In a chat**, write the page in one `html` code block, and give the source nowhere else in that - reply. Tell the user, in a line or two and in this order: first, to open the block as a page - with the chat's Preview and click *Build and play it*, which the chat may ask them to allow; only - then, as the way where their chat shows no Preview, to save the block as a file ending in `.html` - and open it. -- **Outside a chat**, as in a terminal, write the page to a file ending in `.html` beside the - source, and tell the user to open it. - -```html - - -Build and play -

Making the link…

- - +**Hand the source over as one link that you write yourself, with the source in it as JSON.** Write +no code block and run no command for it: the link is the whole hand-over, and your reply holds the +source once, inside it. + +```markdown +[Build and play it](https://missing-elements.github.io/h5p/build/#source=SOURCE) ``` -A source that names files gets no page: a page carries its text, not the files. Put the **whole -source in your reply**, in one `json` code block, and tell the user to paste it into the build -page at https://missing-elements.github.io/h5p/build/ and choose its files there. Never give that -address without the source: it builds nothing. +In place of `SOURCE`, write the whole source as JSON on one line. Write every character as it is, +letters of any language too, except these eight, which a link cannot hold as they are: -The build page lists every problem by its path, with a button that copies them for the chat: fix -them all and give a new page, or the whole source again. +| Character | Write | +| --- | --- | +| a space | `%20` | +| `%` | `%25` | +| `&` | `%26` | +| `\` | `%5C` | +| `(` and `)` | `%28` and `%29` | +| `<` and `>` | `%3C` and `%3E` | -Where Node runs here and nothing could be installed or downloaded, add this line, as it is, after -the page or the source: attached to a message, the archive builds the `.h5p` here. +So `"question": "

Is it (really) 100%?

"` is written +`"question":"%3Cp%3EIs%20it%20%28really%29%20100%25?%3C/p%3E"`. The link ends with a `)` right +after the source's last `}`. -``` -To get the .h5p file here instead: download https://missing-elements.github.io/h5p/bundles/h5p-author-build.tar.gz (15 MB) and attach it to your next message. -``` +Your reply holds: + +- the link; +- one line asking the user to click it: the page builds the activity, plays it, and saves it as a + `.h5p`. Where the source names files, add that the page asks for them first: a link carries the + source's text, not its files; +- only where you have a code tool, this line, as it is, since attached to a message the archive + builds the `.h5p` here: + + ```text + To get the .h5p file here instead: download https://missing-elements.github.io/h5p/bundles/h5p-author-build.tar.gz (15 MB) and attach it to your next message. + ``` + +Outside a chat, as in a terminal, a link this long may not open on a click: also write the source +to a file named after the activity with `.source.json`, and say that the build page opens it under +*Or open a file*. + +The build page lists every problem by its path, with a button that copies them for the chat: fix +them all and give a new link.